In order to update the root port or TT's bandwidth interval table, we will
need to keep track of a list of endpoints, per interval. That way we can
easily know the new largest max packet size when we have to remove an
endpoint.
Add an endpoint list for each root port or TT structure, sorted by
endpoint max packet size. Insert new endpoints into the list such that
the head of the list always has the endpoint with the greatest max packet
size. Only insert endpoints and update the interval table with new
information when those endpoints are periodic.
Make sure to update the number of active TTs when we add or drop periodic
endpoints. A TT is only considered active if it has one or more periodic
endpoints attached (control and bulk are best effort, and counted in the
20% reserved on the high speed bus). If the number of active endpoints
for a TT was zero, and it's now non-zero, increment the number of active
TTs for the rootport. If the number of active endpoints was non-zero, and
it's now zero, decrement the number of active TTs.
We have to be careful when we're checking the bandwidth for a new
configuration/alt setting. If we don't have enough bandwidth, we need to
be able to "roll back" the bandwidth information stored in the endpoint
and the root port/TT interval bandwidth table. We can't just create a
copy of the interval bandwidth table, modify it, and check the bandwidth
with the copy because we have lists of endpoints and entries can't be on
more than one list. Instead, we copy the old endpoint bandwidth
information, and use it to revert the interval table when the bandwidth
check fails.
We don't check the bandwidth after endpoints are dropped from the interval
table when a device is reset or freed after a disconnect, because having
endpoints use less bandwidth should not push the bandwidth usage over the
limits. Besides which, we can't fail a device disconnect.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
In the upcoming patches, we'll use some stored endpoint information to
make software keep track of the worst-case bandwidth schedule. We need to
store several variables associated with each periodic endpoint:
- the type of endpoint
- Max Packet Size
- Mult
- Max ESIT payload
- Max Burst Size (aka number of packets, stored in one-based form)
- the endpoint interval (normalized to powers of 2 microframes)
All this information is available to the hardware, and stored in its
device output context. However, we need to ensure that the new
information is stored before the xHCI driver drops the xhci->lock to wait
on the Configure Endpoint command, so that another driver requesting a
configuration or alt setting change will see the update. The Configure
Endpoint command will never fail on the hardware that needs this software
bandwidth checking (assuming the slot is enabled and the flags are set
properly), so updating the endpoint info before the command completes
should be fine.
Until we add in the bandwidth checking code, just update the endpoint
information after the Configure Endpoint command completes, and after a
Reset Device command completes. Don't bother to clear the endpoint
bandwidth info when a device is being freed, since the xhci_virt_ep is
just going to be freed anyway.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
For upcoming patches, we need to keep information about the bandwidth
domains under the xHCI host. Each root port is a separate primary
bandwidth domain, and each high speed hub's TT (and potentially each port
on a multi-TT hub) is a secondary bandwidth domain.
If the table were in text form, it would look a bit like this:
EP Interval Sum of Number Largest Max Max Packet
of Packets Packet Size Overhead
0 N mps overhead
...
15 N mps overhead
Overhead is the maximum packet overhead (for bit stuffing, CRC, protocol
overhead, etc) for all the endpoints in this interval. Devices with
different speeds have different max packet overhead. For example, if
there is a low speed and a full speed endpoint that both have an interval
of 3, we would use the higher overhead (the low speed overhead). Interval
0 is a bit special, since we really just want to know the sum of the max
ESIT payloads instead of the largest max packet size. That's stored in
the interval0_esit_payload variable. For root ports, we also need to keep
track of the number of active TTs.
For each root port, and each TT under a root port, store some information
about the bandwidth consumption. Dynamically allocate an array of root
port bandwidth information for the number of root ports on the xHCI host.
Each root port stores a list of TTs under the root port. A single TT hub
only has one entry in the list, but a multi-TT hub will have an entry per
port.
When the USB core says that a USB device is a hub, create one or more
entries in the root port TT list for the hub. When a device is deleted,
and it is a hub, search through the root port TT list and delete all
TT entries for the hub. Keep track of which TT entry is associated with a
device under a TT.
LS/FS devices attached directly to the root port will have usb_device->tt
set to the roothub. Ignore that, and treat it like a primary bandwidth
domain, since there isn't really a high speed bus between the roothub and
the host.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Since the xHCI driver now has split USB2/USB3 roothubs, devices under each
roothub can have duplicate "fake" port numbers. For the next set of
patches, we need to keep track of the "real" port number that the xHCI
host uses to index into the port status arrays.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Move the code to check whether we've reached the host controller's limit
on the number of endpoints out of the two conditional statements, to
remove duplicate code.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
The "port" field in xhci_virt_dev stores the port number associated with
one of the two xHCI split roothubs, not the unique port number the xHCI
hardware uses. Since we'll need to store the real hardware port number in
future patches, rename this field to "fake_port".
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Some alternate interface settings have no endpoints associated with them.
This shows up in some USB webcams, particularly the Logitech HD 1080p,
which uses the uvcvideo driver. If a driver switches between two alt
settings with no endpoints, there is no need to issue a configure endpoint
command, because there is no endpoint information to update.
The only time a configure endpoint command with just the add slot flag set
makes sense is when the driver is updating hub characteristics in the slot
context. However, that code never calls xhci_check_bandwidth, so we
should be safe not issuing a command if only the slot context add flag is
set.
Signed-off-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
This patch enables DMA mode1 for the RX path when we know
there won't be any short packets. We check that by looking
into the short_no_ok flag, if it's true we enable mode1, otherwise
we use mode0 to transfer the data.
This will result in a throughput performance gain of around
40% for USB mass-storage/mtp use cases.
[ balbi@ti.com : updated commit log and code comments slightly ]
Signed-off-by: Anand Gadiyar <gadiyar@ti.com>
Signed-off-by: Moiz Sonasath <m-sonasath@ti.com>
Signed-off-by: Vikram Pandita <vikram.pandita@ti.com>
Tested-by: Vikram Pandita <vikram.pandita@ti.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch fixes the compilation brekage which
commits 208466dc ("usb: otg:OMAP4430: Powerdown
the internal PHY when USB is disabled") and
fb91cde4 ("usb: musb: OMAP4430: Power down
the PHY during board init") introduced when
building a OMAP2-only kernel.
LD .tmp_vmlinux1
arch/arm/mach-omap2/built-in.o:(.data+0x7ce0): undefined reference to
+`omap4430_phy_init'
arch/arm/mach-omap2/built-in.o:(.data+0x7ce4): undefined reference to
+`omap4430_phy_exit'
arch/arm/mach-omap2/built-in.o:(.data+0x7ce8): undefined reference to
+`omap4430_phy_power'
arch/arm/mach-omap2/built-in.o:(.data+0x7cec): undefined reference to
+`omap4430_phy_set_clk'
arch/arm/mach-omap2/built-in.o:(.data+0x7cf0): undefined reference to
+`omap4430_phy_suspend'
make: *** [.tmp_vmlinux1] Error 1
Reported-by: Paul Walmsley <paul@pwsan.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
On Audio class, the wLength field of the Setup
packet, contains the data payload size of the
following Data phase. Instead of harcoding values,
use wLength.
This also fixes a bug where Gadget driver had to
receive 3 bytes, but it was queueing a ZLP.
Signed-off-by: Felipe Balbi <balbi@ti.com>
While testing g_audio with HighSpeed UDC on a
FS Hub, we had no configurations to present to
the host. That's because both speeds where
mutually exclusive.
Signed-off-by: Felipe Balbi <balbi@ti.com>
FSG_NUM_BUFFERS is set to 2 as default.
Usually 2 buffers are enough to establish a good buffering pipeline.
The number may be increased in order to compensate a for bursty VFS
behaviour.
Here follows a description of system that may require more than
2 buffers.
* CPU ondemand governor active
* latency cost for wake up and/or frequency change
* DMA for IO
Use case description.
* Data transfer from MMC via VFS to USB.
* DMA shuffles data from MMC and to USB.
* The CPU wakes up every now and then to pass data in and out from VFS,
which cause the bursty VFS behaviour.
Test set up
* Running dd on the host reading from the mass storage device
* cmdline: dd if=/dev/sdb of=/dev/null bs=4k count=$((256*100))
* Caches are dropped on the host and on the device before each run
Measurements on a Snowball board with ondemand_governor active.
FSG_NUM_BUFFERS 2
104857600 bytes (105 MB) copied, 5.62173 s, 18.7 MB/s
104857600 bytes (105 MB) copied, 5.61811 s, 18.7 MB/s
104857600 bytes (105 MB) copied, 5.57817 s, 18.8 MB/s
FSG_NUM_BUFFERS 4
104857600 bytes (105 MB) copied, 5.26839 s, 19.9 MB/s
104857600 bytes (105 MB) copied, 5.2691 s, 19.9 MB/s
104857600 bytes (105 MB) copied, 5.2711 s, 19.9 MB/s
There may not be one optimal number for all boards. This is why
the number is added to Kconfig. If selecting USB_GADGET_DEBUG_FILES
this value may be set by a module parameter as well.
Signed-off-by: Per Forlin <per.forlin@linaro.org>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Acked-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Felipe Balbi <balbi@ti.com>
This patch (as1481) fixes a problem affecting g_file_storage and
g_mass_storage when running at SuperSpeed. The two drivers currently
assume that the bulk-out maxpacket size can evenly divide the SCSI
block size, which is 512 bytes. But SuperSpeed bulk endpoints have a
maxpacket size of 1024, so the assumption is no longer true.
This patch removes that assumption from the drivers, by getting rid of
a small optimization (they try to align VFS reads and writes on page
cache boundaries). If a command's starting logical block address is
512 bytes below the end of a page, it's not okay to issue a USB
command for just those 512 bytes when the maxpacket size is 1024 -- it
would result in either babble (for an OUT transfer) or a short packet
(for an IN transfer).
Also, for backward compatibility, the test for writes extending beyond
the end of the backing storage has to be changed. If the host tries
to do this, we should accept the data that fits in the backing storage
and ignore the rest. Because the storage's end may not align with a
USB packet boundary, this means we may have to accept a USB OUT
transfer that extends beyond the end of the storage and then write out
only the part of the data that fits.
Signed-off-by: Alan Stern <stern@rowland.harvard.edu>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
Now the mass storage driver has fixed logic block size of 512 bytes.
The mass storage gadget read/write bound devices only through VFS, so the
bottom level devices actually are just RAW devices to the driver and connected
PC. As a RAW, hosts can always format, read and write it right in 512 bytes
logic block and don't care about the actual logic block size of devices bound
to the gadget.
But if we want to share the bound block device partition between target board
and PC, in case the logic block size of the bound block device is 4KB, we
execute the following steps:
1. connect a board with mass storage gadget to PC(the board has set one
partition of on-board block device as file name of the mass storage)
2. PC format the mass storage to VFAT by default logic block size and
read/write it
3. disconnect boards from PC
4. target board mount the partition as VFAT
Step 4 will fail since kernel on target thinks the logic block size of the
bound partition as 4KB.
A typical error is "FAT: logical sector size too small for device (logical
sector size = 512)"
If we execute opposite steps:
1. format the partition to VFAT on target board and read/write this partition
2. connect the board to Windows PC as usb mass storage gadget, windows will
think the disk is not formatted
So the conclusion is that only as a gadget, the mass storage driver has no any
problem. But being shared VFAT or other filesystem on PC and target board, it
will fail.
This patch adapts logic block size to bound block devices and fix the issue.
Cc: Michal Nazarewicz <mina86@mina86.com>
Acked-by: Alan Stern <stern@rowland.harvard.edu>
Signed-off-by: Peiyu Li <peiyu.li@csr.com>
Signed-off-by: Xianglong Du <xianglong.du@csr.com>
Signed-off-by: Huayi Li <huayi.li@csr.com>
Signed-off-by: Barry Song <Baohua.Song@csr.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
It's useful to know which states core is going
through, as it might help us figure out misbehavior
on specific link states.
Signed-off-by: Felipe Balbi <balbi@ti.com>
This flag will tell us which direction we're
expecting on the next (data or status) phase.
It will help us catching errors of host going
crazy and requesting data of the wrong direction.
Signed-off-by: Felipe Balbi <balbi@ti.com>
if req->dma isn't DMA_ADDR_INVALID it means gadget driver
mapped the request or allocated from coherent, so it's
unnecessary to do anything.
Signed-off-by: Felipe Balbi <balbi@ti.com>
If the gadget drivers sends a ZLP we are trying to map this this request
which does not work on all implementations. So we simply skip mapping
it.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
dwc3-wrapper can be used by any other wrapper,
using dwc3-omap makes it clear that we're running
on OMAP SoC.
Signed-off-by: Felipe Balbi <balbi@ti.com>
The OMAP wrapper allows us to either control internal
OTG signals via SW or HW. Different boards might wish
to use one or the other mode of operation. Let's have
have that information passed via platform_data for now.
After DT conversion is finished for OMAP, we can easily
convert this to a DT attribute.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We need to have actual HW in order to implement
and test that part of the code anyway. Until then
it's best to remove it.
Signed-off-by: Felipe Balbi <balbi@ti.com>
this field will hold the next expected event.
In certain cases, host might fall into some error
condition and ask from us the wrong Control phase.
On such situations, we should stall and restart.
Signed-off-by: Felipe Balbi <balbi@ti.com>
Whenever we issue a Set Stall command on EP0,
the state machine will be restarted and Stall
is cleared automatically, when core receives
the next SETUP packet.
There's no need to track that EP0_STALL state.
Signed-off-by: Felipe Balbi <balbi@ti.com>
when we're going to issue Set Stall command,
we should clear DWC3_EP_STALL flag, but also
we should clear BUSY, HALTED and all others.
Signed-off-by: Felipe Balbi <balbi@ti.com>
There are two spots where we wait until the HW finishes processing a
certain command. Initially we had a few problems and we used 500ms as a
limit to be on a the safe side. Paul Zimmerman mentioned this is little too
much. After a debugging session, we noticed that we hardly ever go over 20us
and didn't pass 30usec so far. Using mdelay() seems way overloaded.
Giving the current numbers 500usec as the upper limit is more than enough.
Should it ever timeout then something is definitely wrong.
While here, also replace the type with u32 since long does not really
fit here.
Cc: Paul Zimmerman <paul.zimmerman@synopsys.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
- since a while we are disabling an endpoint and purging every requests on
RESET and DISCONNECT which leads to a warning since the endpoint was
disabled twice (once by the UDC, and second time by the gadget). I
think UDC should nuke all requests because all those requests
become invalid. It's gadget driver's responsability, though, to disable
its used endpoints. This is done by merging dwc3_stop_active_transfer()
and dwc3_gadget_nuke_reqs() into dwc3_remove_requests().
- dwc3_stop_active_transfer() is now no longer called unconditionaly.
This has the advantage that it is always called to disable an active
transfer which means if res_trans_idx 0 than something went wrong and
it is an error condition because we can't clean up the requests.
- Remove the DWC3_EP_WILL_SHUTDOWN which was introduced while
introducing the command complete part for dequeue. All requests on
req_queued list should be removed during the dwc3_cleanup_done_reqs()
callback so there is no reason to go through the list again.
We consider it an error condition if requests are still on this
list since we never queue TRB without LST=1 (the last requests has
always LST=1, there are no requests with LST=0 behind it).
[ balbi@ti.com : reworked commit log a bit, made patch apply ]
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
We read the DWC3_GSNPSID register to make sure we got the correct
register offset passed. One of the recent commits moved the soft reset
before this so in case of the wrong offset we end up with "reset timed
out". This patch moves the "id" check before the reset again.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
There are some issues around for enabling/disabling this mode and
handling it. It does not work perfectly (yet). However we have a few
gadgets tested successfuly so far. That means we are quite confident
that we won't need this in near future.
So I'm for removing it and bringing a working version back once there is
a need for it.
Thanks to Dan Carpenter who spotted the wrong memory handling here.
[ balbi@ti.com : made it actually apply ]
Cc: Dan Carpenter <error27@gmail.com>
Cc: wharms@bfs.de
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Felipe Balbi <balbi@ti.com>
The DesignWare USB3 core tells us which phase
of a control transfer should be started, it
also tells us which physical endpoint needs
that transfer.
With these two informations, we have all we
need to simply EP0 handling quite a lot and
get rid rid of the SW state machine tracking
ep0 states.
For achieving this perfectly, we needed to
add support for situations where we get
XferNotReady while endpoint is still busy
and XferNotReady while gadget driver still
hasn't queued a request.
Signed-off-by: Felipe Balbi <balbi@ti.com>
In case we have transfers which aren't aligned
to wMaxPacketSize, we need to be careful with
how we start the transfer with the HW. OUT
transfers _must_ be aligned with wMaxPacketSize
and in order to guarantee that, we use a bounce
buffer.
Signed-off-by: Felipe Balbi <balbi@ti.com>
This core cannot handle OUT transfers which aren't
aligned to wMaxPacketSize, but that can happen at
least on control endpoint with the USB Audio Class.
This patch adds a bounce buffer to be used on the
case of a non-aligned ep0out request is queued.
Signed-off-by: Felipe Balbi <balbi@ti.com>
The status field of the Transfer Not Read event
is different on Control Endpoints. On this patch
we are just adding the defines to be used on a
later patch which will re-work the control endpoint
handling.
Signed-off-by: Felipe Balbi <balbi@ti.com>
the previous message had too little meaning. Make
it more human readable and use the macro we already
had for extracting the command completion status out
of DEPCMDn register.
Signed-off-by: Felipe Balbi <balbi@ti.com>
if we don't set DMA address to invalid when unmapping,
we might fall in a situation where request buffer
can't be mapped to DMA again.
Signed-off-by: Felipe Balbi <balbi@ti.com>
We may as well fix this potential leak so we don't have to listen to
the static checkers complain.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Felipe Balbi <balbi@ti.com>
